Splitting processor and target in BSPs

Splitting processor and target in BSPs

[Chris Tapp]

How should meta data be structured so that a layer can support a set  
of systems using a set of processors?

For example, many of the 'eBox' systems use variants of the Vortex86  
SoC. So, a set of machine files are needed for these (e.g. ebox-3300,  
ebox-3500mx, etc.).

These have different peripherals available (e.g. some have serial,  
some don't) and use different SoC variants with different cpu, sound,  
etc. It would therefore make sense for the machine configuration to  
inherit the SoC attributes (for the common features) and add (or  
remove) machine specific attributes (e.g. serial) to these. This can  
be done by putting the SoC bits in to an include.

However, kernel configuration becomes a little bit more complicated as  
this is done by machine name. A kernel recipe will be needed for each  
machine (e.g. for the different sound drivers), but I can't work out  
how to do this using a base configuration for the SoCs that are shared  
and then adding machine specific parts. I can do it using (for  
example) a .defconfig for each machine, but that would require updates  
to multiple files to change the SoC configuration.

I guess what I'm really asking is, is it possible to have a base CPU  
configuration and add a machine configuration to this ?

I've recently seen discussion of .cfg kernel fragment files. Are these  
what I should be looking at? Are these available in the releases or  
only in the development branch?

Chris Tapp

opensource@keylevel.com
www.keylevel.com

Re: Splitting processor and target in BSPs

[Bruce Ashfield]

On 11-09-02 03:26 AM, Chris Tapp wrote:
> How should meta data be structured so that a layer can support a set of
> systems using a set of processors?
>
> For example, many of the 'eBox' systems use variants of the Vortex86
> SoC. So, a set of machine files are needed for these (e.g. ebox-3300,
> ebox-3500mx, etc.).
>
> These have different peripherals available (e.g. some have serial, some
> don't) and use different SoC variants with different cpu, sound, etc. It
> would therefore make sense for the machine configuration to inherit the
> SoC attributes (for the common features) and add (or remove) machine
> specific attributes (e.g. serial) to these. This can be done by putting
> the SoC bits in to an include.
>
> However, kernel configuration becomes a little bit more complicated as
> this is done by machine name. A kernel recipe will be needed for each
> machine (e.g. for the different sound drivers), but I can't work out how
> to do this using a base configuration for the SoCs that are shared and
> then adding machine specific parts. I can do it using (for example) a
> .defconfig for each machine, but that would require updates to multiple
> files to change the SoC configuration.
>
> I guess what I'm really asking is, is it possible to have a base CPU
> configuration and add a machine configuration to this ?
>
> I've recently seen discussion of .cfg kernel fragment files. Are these
> what I should be looking at? Are these available in the releases or only
> in the development branch?

What you've described is one of the primary reasons that the
configuration fragments exist :) You define a common set of
config fragments, extend and then select what you want.

These are available in all the releases, but the capabilities
are evolving and I merge more changes into the yocto bindings
for manipulating the configuration fragments outside of a kernel
tree proper.

If you have your own kernel repository based on linux-yocto, you
could implement any of this inside the tree. If you are re-using
existing machines/branches in the linux-yocto tree, you can also
do what you want, but have to do it via a collection of configuration
fragments that are appended to the SRC_URI based on the machine
you are building,.

Cheers,

Bruce

>
> Chris Tapp
>
> opensource@keylevel.com
> www.keylevel.com
>
>
>
> _______________________________________________
> yocto mailing list
> yocto@yoctoproject.org
> https://lists.yoctoproject.org/listinfo/yocto

Re: Splitting processor and target in BSPs

[McClintock Matthew-B29882]

On Fri, Sep 2, 2011 at 2:26 AM, Chris Tapp <opensource@keylevel.com> wrote:
> However, kernel configuration becomes a little bit more complicated as this
> is done by machine name. A kernel recipe will be needed for each machine
> (e.g. for the different sound drivers), but I can't work out how to do this
> using a base configuration for the SoCs that are shared and then adding
> machine specific parts. I can do it using (for example) a .defconfig for
> each machine, but that would require updates to multiple files to change the
> SoC configuration.
>
> I guess what I'm really asking is, is it possible to have a base CPU
> configuration and add a machine configuration to this ?
>
> I've recently seen discussion of .cfg kernel fragment files. Are these what
> I should be looking at? Are these available in the releases or only in the
> development branch?

I think this is a good solution for you. It's a little confusing to
find were this work is but you can look in the linux-yocto kernel tree
- then the actual cfg fragment stuff on different branches.

http://git.yoctoproject.org/cgit/cgit.cgi/linux-yocto-3.0/

meta branch contains meta/ which contains a lot of the cfg fragments,
and patches, etc. There is also a document in the git repo as well
which is worth a look.

-M

Re: Splitting processor and target in BSPs

[Tom Zanussi]

On Fri, 2011-09-02 at 00:26 -0700, Chris Tapp wrote:
> How should meta data be structured so that a layer can support a set  
> of systems using a set of processors?
> 
> For example, many of the 'eBox' systems use variants of the Vortex86  
> SoC. So, a set of machine files are needed for these (e.g. ebox-3300,  
> ebox-3500mx, etc.).
> 
> These have different peripherals available (e.g. some have serial,  
> some don't) and use different SoC variants with different cpu, sound,  
> etc. It would therefore make sense for the machine configuration to  
> inherit the SoC attributes (for the common features) and add (or  
> remove) machine specific attributes (e.g. serial) to these. This can  
> be done by putting the SoC bits in to an include.
> 
> However, kernel configuration becomes a little bit more complicated as  
> this is done by machine name. A kernel recipe will be needed for each  
> machine (e.g. for the different sound drivers), but I can't work out  
> how to do this using a base configuration for the SoCs that are shared  
> and then adding machine specific parts. I can do it using (for  
> example) a .defconfig for each machine, but that would require updates  
> to multiple files to change the SoC configuration.
> 
> I guess what I'm really asking is, is it possible to have a base CPU  
> configuration and add a machine configuration to this ?
> 
> I've recently seen discussion of .cfg kernel fragment files. Are these  
> what I should be looking at? Are these available in the releases or  
> only in the development branch?
> 

You might also be able to use 'KERNEL_FEATURES' to help manage the .cfg
fragments by machine.  For example, in the kernel recipes you can see:

meta/recipes-kernel/linux/linux-yocto_3.0.bb:

    # Functionality flags
    KERNEL_FEATURES="features/netfilter"
    KERNEL_FEATURES_append=" features/taskstats"
    KERNEL_FEATURES_append_qemux86=" cfg/sound"
    KERNEL_FEATURES_append_qemux86-64=" cfg/sound"

which map to features in the yocto kernel meta branch.

And in the kernel .bbappends in the kernel dev layer some more
explanation:

poky-extras/meta-kernel-dev/recipes-kernel/linux/linux-yocto_3.0.bbappend:

    # KERNEL_FEATURES are features to be added to the kernel, and must
    # point to configurations stored on the 'meta' branch of the kernel
    # that is being built.
    # KERNEL_FEATURES ?= <FOO>

Tom

> Chris Tapp
> 
> opensource@keylevel.com
> www.keylevel.com
> 
> 
> 
> _______________________________________________
> yocto mailing list
> yocto@yoctoproject.org
> https://lists.yoctoproject.org/listinfo/yocto

Re: Splitting processor and target in BSPs

[Chris Tapp]

On 2 Sep 2011, at 16:49, McClintock Matthew-B29882 wrote:

> On Fri, Sep 2, 2011 at 2:26 AM, Chris Tapp <opensource@keylevel.com>  
> wrote:
>> However, kernel configuration becomes a little bit more complicated  
>> as this
>> is done by machine name. A kernel recipe will be needed for each  
>> machine
>> (e.g. for the different sound drivers), but I can't work out how to  
>> do this
>> using a base configuration for the SoCs that are shared and then  
>> adding
>> machine specific parts. I can do it using (for example)  
>> a .defconfig for
>> each machine, but that would require updates to multiple files to  
>> change the
>> SoC configuration.
>>
>> I guess what I'm really asking is, is it possible to have a base CPU
>> configuration and add a machine configuration to this ?
>>
>> I've recently seen discussion of .cfg kernel fragment files. Are  
>> these what
>> I should be looking at? Are these available in the releases or only  
>> in the
>> development branch?
>
> I think this is a good solution for you. It's a little confusing to
> find were this work is but you can look in the linux-yocto kernel tree
> - then the actual cfg fragment stuff on different branches.
> http://git.yoctoproject.org/cgit/cgit.cgi/linux-yocto-3.0/
>
> meta branch contains meta/ which contains a lot of the cfg fragments,
> and patches, etc.

Thanks, I can see what they look like now ;-)

What are 'the rules' for cfg files? I.e.:

1) Does adding #CONFIG_this_bit not set
disable a node and its dependencies? For example, if I turn off usb  
serial support do the usb serial drivers get disabled as well or do I  
have to do this manually?

2) Similarly, if a node is enabled what happens to the dependent  
nodes? Do they all need to be covered, or only the ones that need to  
be enabled?

I've also noticed that there are scc files. These seem to pull  
features in to sets? How do these work?

meta/cfg/kernel-cache/bsp/common-pc/hardware.cfg contains

CONFIG_ATH5K
CONFIG_ATH_COMMON
...
CONFIG_SERIAL_8250_CONSOLE

It seems as if this file does something else, as these don't look like  
valid kernel configuration options.

> There is also a document in the git repo as well which is worth a  
> look.

I couldn't see one, but there's a lot in there ;-)

Sorry for all the questions, but there's a lot to understand for  
someone new :-)


Chris Tapp

opensource@keylevel.com
www.keylevel.com

Re: Splitting processor and target in BSPs

[Bruce Ashfield]

On 11-09-03 7:32 AM, Chris Tapp wrote:
> On 2 Sep 2011, at 16:49, McClintock Matthew-B29882 wrote:
>
>> On Fri, Sep 2, 2011 at 2:26 AM, Chris Tapp <opensource@keylevel.com>
>> wrote:
>>> However, kernel configuration becomes a little bit more complicated
>>> as this
>>> is done by machine name. A kernel recipe will be needed for each machine
>>> (e.g. for the different sound drivers), but I can't work out how to
>>> do this
>>> using a base configuration for the SoCs that are shared and then adding
>>> machine specific parts. I can do it using (for example) a .defconfig for
>>> each machine, but that would require updates to multiple files to
>>> change the
>>> SoC configuration.
>>>
>>> I guess what I'm really asking is, is it possible to have a base CPU
>>> configuration and add a machine configuration to this ?
>>>
>>> I've recently seen discussion of .cfg kernel fragment files. Are
>>> these what
>>> I should be looking at? Are these available in the releases or only
>>> in the
>>> development branch?
>>
>> I think this is a good solution for you. It's a little confusing to
>> find were this work is but you can look in the linux-yocto kernel tree
>> - then the actual cfg fragment stuff on different branches.
>> http://git.yoctoproject.org/cgit/cgit.cgi/linux-yocto-3.0/
>>
>> meta branch contains meta/ which contains a lot of the cfg fragments,
>> and patches, etc.
>
> Thanks, I can see what they look like now ;-)
>
> What are 'the rules' for cfg files? I.e.:
>
> 1) Does adding #CONFIG_this_bit not set
> disable a node and its dependencies? For example, if I turn off usb
> serial support do the usb serial drivers get disabled as well or do I
> have to do this manually?

There's no extra processing on top of what LKC gets you, so yes,
that's how it would work. If you disable a top level option in
your fragment, lkc will then disable everything that depends on it.

>
> 2) Similarly, if a node is enabled what happens to the dependent nodes?
> Do they all need to be covered, or only the ones that need to be enabled?

Same as #1. If you select FOO, you can now select anything that
depends on FOO. If FOO has its own "select BAR" statements in it,
they'll be enabled in the final .config.

So just cover what needs to be enabled.

>
> I've also noticed that there are scc files. These seem to pull features
> in to sets? How do these work?

It's best to check the yocto kernel manual for some desriptions on
this, rather than me writing it all here. I also have an updated set
of docs for 1.1 that I can send along shortly.

The .scc files are feature descriptions that cover patches (if you
have any), git operations (if you have any) and configuration. They
allow a feature and configuration to be kept together and associated.
Once in a .scc file, that feature + config can be re-use and imported
by other features.

>
> meta/cfg/kernel-cache/bsp/common-pc/hardware.cfg contains
>
> CONFIG_ATH5K
> CONFIG_ATH_COMMON
> ...
> CONFIG_SERIAL_8250_CONSOLE
>
> It seems as if this file does something else, as these don't look like
> valid kernel configuration options.

Covered in some of the docs, but there is a kernel configuation audit that
runs at the end. It is intended to let the user know if something they
didn't expect happened. i.e.:

   - you had CONFIG_FOO=y in your fragment, but the final .config didn't
     or had it as another value.
   - high level configuration fragments are supposed to set non-hardware
     options that are a 'policy' for the system. And BSPs should be setting
     hardware options. These are the "kconf hardware <foo>.cfg" and
     "kconf non-hardware <bar>.cfg" that you see in the files. If a BSP
     option overrides a system policy (and hence may behave differently
     in an important way), you get a warning.
   - if an option is repeated you get a warning
   - if a BSP sets a non-hardware option you get a warning. The config
     audit has 'buckets' for options that are hardware and non-hardware,
     so it can emit this warning. But that's not perfect. If a BSP developer
     says 'hey, this IS hardware for my board', you can list those options
     in "hardware.cfg" in your BSP dir and the audit system will leave
     these options out of testing.

>
>> There is also a document in the git repo as well which is worth a look.
>
> I couldn't see one, but there's a lot in there ;-)
>
> Sorry for all the questions, but there's a lot to understand for someone
> new :-)

Check the docs of the yocto twiki page for some theory and examples. And
I'm happy to fill in any blanks (since those would be doc gaps, and I'll
write something up).

Bruce

>
>
> Chris Tapp
>
> opensource@keylevel.com
> www.keylevel.com
>
>
>
> _______________________________________________
> yocto mailing list
> yocto@yoctoproject.org
> https://lists.yoctoproject.org/listinfo/yocto